DevourerOfTime

Oh wow, I just so happened to be playing this game at the same time as the game club here.

I have been sucked into this game over the last week in a way only Civilization games have done in the past, even with the lack of depth and challenge. I'm onto my third playthrough and am close to finishing all the achievements, something I rarely care about other than as an excuse to play a game for more time. We'll see if I actually finish it as the last achievement that will be left is just an enormous time waster (100km of roads is like an order of magnitude too many roads for even multiple playthroughs).

Still, I'm really enjoying my time with it. It's time waster junk food, but its good time waster junk food. It would have made my game of the year list last year if I played it before the year ended.

@thepanzini: The problem with any of these is that it doesn't show the training data. Nor is there any way to prove that all of the training data was used with the permission of the copyright holder of those images/models/whatever.

You can say "ah, I used Midjourney for this", but that's just differing the responsibility onto someone else. Midjourney isn't open about it's training data. It could be using every copyrighted image in history and there is no way to prove it isn't using copyrighted material.

In fact, we have proof of the opposite.

So now that we know this, any material made with Midjourney is tainted. Until they purge that material from its training data, anything made with Midjourney was made while that was in the training data unknowingly was using copyrighted material without the permission of the artist or the copyright holder of that art.

And that's just one (1) copyrighted material holder. Like, if you get a generative A.I. to make any copyrighted character, it requires you to have training data from that copyright. You can't make a picture of Goku unless the A.I. knows what a Goku is and is trained to know what Goku looks like, which requires an image of Goku. And a company owns the image of what Goku looks like. That's how this works.

I know I can post a picture of Goku that I've drawn in this post, but I'm not selling this post, so Toei Animation graciously doesn't smite me from space.

I'm not saying that point just on an ethical standpoint. As I point out when I wrote the concept page, this is all super legally shakey.

Generative A.I. is going to have a place in games because company heads want there to be, but it's going to be a bumpy ride for anyone who adopts this early and thinks its completely on the level.

@thepanzini:While the output can seem similar, the difference is in how the sausage is made.

No Man's Sky has a procedural generation system that takes pre-made assets made by a person and stitches them together in new ways. This is true for the plants, animals, terrain, planets, ships, aliens, space stations, etc. etc. All of those assets were made by a person, even if that combination was never what was intended. It's essentially just a LOT of smart coding and efficient use of created assets that is then thrown into a random number generator. The seed of that RNG tells the algorithm what to make where and in what way.

Generative A.I. doesn't work like that. It is an algorithm someone created, yes, but not to specifically to create a No Man's Sky like galaxy. Now, again, I'm no Machine Learning expert, but the best, simple description I can give of what Generative A.I. is this: Generative A.I. is a complex mix of pattern recognition in data (like what objects exist in a picture), natural language interpretation (giving the program plain old English and it converting that into data it can understand and use), and repurposing data it has analyzed and accumulated into an output that fits the parameters of its input.

So... what does that mean? It means that for a Generative A.I. to work, you have to feed a LOT of examples into it, all labelled with different information about what is in the input (ex. this image is of a band playing on a stage). Once the Generative A.I. has enough data points of what a band looks like playing on a stage, it will start being able to repurpose data it has already taken in and spit out a new output on what an image of a band playing on a stage looks like. It's not creating something new, just stitching colour values, patterns, and details together to be a certain way that is in line with what it has been told what a band being on stage looks like.

Now with this, Generative A.I. and Procedural Generation might still sound the same, but there are four crucial changes to how Generative A.I. works.

1) Procedural Generation uses input that is all made by the developer, using exactly the media and assets that were made to work with it.

Generative A.I., however, for it to spit out a "new" image of a band, it has to have a VERY LARGE amount of images that are of bands on stages that have been labelled and processed. Like WAY MORE than you even think. Way more than any one person could take photos of or draw images of in their life. Where do you get those images from then? It has to be sourced from somewhere. And very few people have given permissions for companies to use their work, regardless of the medium, for use to train these neural networks.

So every single A.I. neural network has been trained off of information and media they have either been scraped from sites that don't even know their work is being fed into an A.I. OR it's a company like Google, Microsoft, Apple, or Meta that have large amounts of data on their servers that they can use without informing people (or at least informing them buried deep in a 200 page terms of service document).

2) In Procedural Generation, the output is curated by the developer by honing the algorithms it uses to ensure the level, animal, whatever is how they want the game to be.

The A.I. has no control or filter over what it outputs. It just sees data. Data that is interpreted one way from input, more data from a prompt from a user, and output that fits the data it has. And there are very few nobs that one can turn to filter this.

Let's say you used A.I. to make Spelunky levels based on existing Spelunky levels. No matter how much valid, awesome, procedurally generated Spelunky levels you feed into the algorithm, it is still going to produce some levels with, say, no way to reach the exit. Or a drop that you can't fall down without hurting yourself. Or an enemy that shouldn't exist in one level being in another. Or lava right above the entrance.

Because the A.I. doesn't have rules to make a thing. It just is given data and output based on what data it has been given before. Sure, none of the levels it had been given had a section where it was impossible to get down without dying, but it doesn't look for that. It just looks at the data and rearranges it to look like data it has seen before.

To give the A.I. all of the data it needs to perfectly make Spelunky levels, you have to feed into it, with each of those levels, all of the information that you already have procedural generation checking for. On top of all those completed levels you have already generated using those algorithms. And all of the assets to make them look authentic.

3) Procedural Generation is stable. Meaning that, except for bugs or oversights in its creation, procedurally generation will always produce output that it is expected to make.

Generative A.I. is constantly changing with each piece of input it is fed and there is no way to reign in if those changes are *bad*.

We've seen this already with text-based Generative A.I., over time losing their ability to do basic arithmetic (you know, the easiest thing a computer can do). And we're going to see this with image-based Generative A.I. very soon too, with artists fighting back against it with A.I. poisoning tools like Nightshade for when their art is inevitably stolen from the internet by their scraping algorithms gathering input to feed their neural network.

And while Nightshade is a malicious act against neural networks (but not unwarranted imo), such acts are, again, not necessary for the A.I. to deteriorate. Just give enough humans the ability to chat with it for a long enough time and the work will be done. Because if you were able to make an A.I. that had checks in place that was good enough to distinguish all data as valid or invalid from the human race, you wouldn't need the A.I. part of it. You would just have just manually programmed something so complex that it has omnipotence.

4) Procedural Generation doesn't fucking steal anything.

Destructoid posted a pretty level-headed take on the whole thing.

https://www.destructoid.com/best-places-to-respawn-early-on-in-palworld/

Yeah this is why I posted here about it's specific inclusion. I am not super well equipped to point out what is Generative A.I. trained on very specific copyrighted data vs. plain old stealing of assets or at least being "heavily inspired by". I really only have my understanding of Machine Learning (which is only basically what I got from my degree and not working in the field everyday or anything).

So I wanted to post here to kind of see what people thought. I should have been more clear on that.

With going over the evidence and with a day to think on it, I think I may have jumped the gun a bit. So I will be removing it for now.

Still, I will in no way be surprised if it ends up that the game was made using Generative A.I. trained on Pokemon designs. I will keep posting about it here if more evidence or news stories come to light (which probably won't be until the week starts).

Based on this article on Video Games Chronicle, I've added Palworld to this concept.

https://www.videogameschronicle.com/news/palworld-embroiled-in-ai-and-pokemon-plagiarism-controversy/

This is developing and I'll update with any confirmations or any other new articles, but evidence is basically flooding in on social media.

I don't think the fight's over against Generative A.I. being used widespread in games

There are still a LOT of legal disputes to figure out about it, not to mention countries that might regulate against it.

Also The Finals is such a weird one for people to be excited about because MAN is that game ugly.

Forgot to add my Other vote, but it was just going to be:

"The Entire Video Game Industry for collectively laying off over 9000 people."

The most anger I've ever received from people about a take of mine was that Super Smash Bros. Brawl was a better competitive game than Super Smash Bros. Melee.

And yes, I both participated in and ran tournaments in both games. In fact, the last time Brawl had a tournament in town I swept the tournament.

But this take gets under the skin of soooo many people. And they are never really equipped to defend it because Melee is this golden perfect child and Brawl is the kid in the class everyone picks on because it is what everyone does.

But I feel like my critiques of Melee are valid.

• It focuses far too much a ludicrous execution barrier that is only fun for the people with the dexterity and willingness to forfeit their long term wrist and hand health to compete. No other game, fighting game or platform fighter, makes you risk that much just to play it.
• Its viable characters on the tier list is small, even given its roster size. What's more is that playstyle diversity is minimal. Yes, even including Jigglypuff. Comparitively, Brawl has a huge mid tier of viable options (like my main character, Kirby), all with very different playstyles that are viable.
• The above also means counterpicks are basically nonexistent in the meta. In Brawl, counterpicks are a lot bigger part of the meta game, as I had 4-5 characters I could bust out for a match if I felt they performed better. Many of my friends who competed did as well.
• Stage diversity is frankly terrible. I mean, it is in Brawl too, but come on, its even worse in Melee.
• Not only is the execution barrier tough to overcome to be competitive, even just to start playing the game, it is a daunting feat, one that might be the highest in platform fighters or fighting games in general.
• My old man back doesn't want to carry a CRT around to tournaments ever again.